A bioinspired hierarchical reinforcement learning architecture for modeling learning of multiple skills with continuous state and actions

Organisms, and especially primates, are able to learn several skills while avoiding catastrophic interference and enhancing generalisation. This paper proposes a novel hierarchical reinforcement learning (RL) architecture with a number of features that make it suitable to investigate such phenomena. The proposed system combines the mixture of experts architecture with the neural-network actor-critic architecture trained with the TD() reinforcement learning algorithm. In particular, responsibility signals provided by two gating networks (one for the actor and one for the critic) are used both to weight the outputs of the respective multiple (expert) controllers and to modulate their learning. The system is tested with a simulated dynamic 2D robotic arm that autonomously learns to reach a target in (up to) three different conditions. The results show that the system is able to appropriately allocate experts to tasks on the basis of the differences and similarities among the required sensorimotor mappings.